[hts-users:00887] Re: Training by imposing decision trees

[Simon King <Simon.King@xxxxxxxx>]

Heiga ZEN (Byung Ha CHUN) wrote:
> In this case parameters of generated MMF are equal to those of MMF 
> used to construct decision trees.
> So you should re-estimate model parameters 

Good point - I hadn't considered the case where the decision tree being 
imposed might have been trained on different data, I just thought about 
the case where it matched the data - although the case of differing data 
is obviously more likely and more interesting.

> However, it may cause some segmentation errors at the first iteration 
> of HERest if you use different training data to construct these two MMFs.

I think HERest offers a solution to this, called "2-model re-estimation" 
in which one model (the one trained on the target data, but before it is 
reclustered) is used to compute the forward-backward probabilities and 
these are then used to update the parameters of a second model (the 
re-clustered one). The tying schemes in the two models do not have to 
match. See the HTK book for details.


Simon